Artificial intelligence helps discover novel antibiotic candidates

Mondo Health Updated on 2024-01-31

Xinhua News Agency, Beijing, Jan. 1 (Xinhua) -- The Massachusetts Institute of Technology (MIT) recently issued a press release saying that an international team of researchers from the university used artificial intelligence deep learning models to discover new compounds that can be infected by drug-resistant bacteria. These compounds have the potential to be novel antibiotic drugs.

Researchers at institutions such as MIT and Harvard University first tested about 3The antimicrobial activity of 90,000 compounds against methicillin-resistant Staphylococcus aureus was used, and the test data and information such as the chemical structure of these compounds were used to train the deep learning model. The research team used an algorithm called Monte Carlo tree search to allow the model to determine not only the antimicrobial activity of each molecule, but also which chemical substructures of the molecule could cause that activity. To further narrow down the drug candidates, they also trained three other deep learning models to determine whether the compounds were toxic to three types of human cells.

Next, the researchers used the collection of the above models** to determine the antimicrobial activity and cytotoxicity of about 12 million compounds and determined that five classes of compounds had antimicrobial activity against methicillin-resistant Staphylococcus aureus. They selected 280 compounds in petri dishes to test against methicillin-resistant Staphylococcus aureus, and finally selected two candidate antibiotics belonging to the same class of compounds. Experimental results in mice showed that both compounds were effective against **methicillin-resistant Staphylococcus aureus infection.

A key innovation of the new study, the researchers say, is that it opens up the "black box" of such deep learning models and figures out what kind of information the model uses to ** antibiotic potency, which will help researchers design drug candidates that work better than the compounds identified by the models. In the future, they will analyze the chemistry and potential clinical use of the two compounds in more detail, and use such deep learning models to design more drug candidates and find compounds that can kill other bacteria.

Related** has been published in the new issue of the British journal Nature.

Related Pages